Re: Fwd: Cluster "stuck" in "not accepting commands to avoid wraparound data loss"

On Thu, Dec 17, 2015 at 1:20 PM, Andres Freund <andres@anarazel.de> wrote:
> On 2015-12-17 13:08:15 -0500, Robert Haas wrote:
>> On Thu, Dec 17, 2015 at 12:14 PM, Andres Freund <andres@anarazel.de> wrote:
>> > On 2015-12-17 09:04:25 -0800, Jeff Janes wrote:
>> >> > But I'm somewhat confused what this has to do with Andres's report.
>> >>
>> >> Doesn't it explain the exact situation he is in, where the oldest
>> >> database is 200 million, but the cluster as a whole is 2 billion?
>> >
>> > There were no crashes, so no, I don't think so.
>>
>> Backing up a step, do we think that the fact that this was running in
>> a shell rather than a screen is relevant somehow?  Or did something
>> happen to this particular cluster totally unrelated to that?
>
> I reran the whole thing on a separate, but very similar, VM. Just
> checked. Same thing happened. This time I have log files and
> everything. No time to investigate right now, but it's reproducible if
> you accept running tests for a week or so.

I don't think I'm going to speculate further until you have time to
investigate more.  It seems clear that autovacuum is going wrong
somehow, but it's extremely unclear why.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company